Hi, > whoops, that results in a matrix with 10^8 (=10k*10k) rows and > columns - just too big for GPU RAM. > > > Is it really that big? My GPU has 3GB of RAM, I can't believe 10^8 > matrix entries is too much for it.
Just compute the memory consumption: The matrix has 10^8 rows, with 5 entries per row. Each element takes 8 bytes of data in double precision, so we need 4*10^9 bytes. In other words: 4 GB ;-) > -Sparse benchmark still reads vector data from testdata files > > ( result65025.txt ). Is it supposed to be this way? > > > No, you can discard this, it's not necessary since it is not a test. > > > But the sparse benchmark heavily depends on this vector data. It is read > into the ublas_vec1 and ublas_vec2 variables, which are used throughout > the entire benchmark. I don't see how it can be discarded. The benchmark only 'depends' on the nonzero pattern of the sparse matrix. It is not at all important which numerical values are in the vectors and the nonzeros of the sparse matrix. If there are any checks for correct results in there, just get rid of them. > BLAS 3 only runs with dense matrices, so there is no need to run any > MatrixMarket stuff to get performance numbers. Just use the > MatrixMarket data for the sparse matrix benchmarks. > > > Alright. So basically, when one hits the download and run button, only > the sparse benchmark should be run with the selected matrix. And the > result should be shown where exactly? The basic and/or expert benchmark > view? Within the MatrixMarket browser module? Both in the views and the MM? At the very least in the basic view, where you just run the set of sparse matrix tests on them (i.e. compare the matrix-vector product performance for different storage formats) > While we are at it: Could you please think about how to incorporate > OpenMP into the GUI? Basically a user should have the following choices: > - Run with OpenMP or OpenCL > - If with OpenCL: Select device on which to run > > This parameter should be selected from the basic view already, just > like the precision. The changes to the existing code are minimal, > one only needs to provide the right viennacl::context() to the > matrix and vector contructors. > > Sure, I'll see what can be done. Gonna need some detailed info on how to > select devices & platforms for the benchmark. As far as I can see, > there's only #ifdef VIENNACL_WITH_OPENCL checking for OpenCL. examples/benchmarks/blas3.cpp has some code for device selections. Other than that, have a look at viennacl::context() and how this is specified in e.g. the vector<> and vector_base<> class. Best regards, Karli ------------------------------------------------------------------------------ Infragistics Professional Build stunning WinForms apps today! Reboot your WinForms applications with our WinForms controls. Build a bridge from your legacy apps to the future. http://pubads.g.doubleclick.net/gampad/clk?id=153845071&iu=/4140/ostg.clktrk _______________________________________________ ViennaCL-devel mailing list ViennaCL-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/viennacl-devel
